I do not come from a technical background. What I had was years of sitting across the table from clients—and that turned out to be exactly what this problem needed.

A bank’s fraud team was drowning in alerts. Volume was growing faster than headcount, while every investigated case carried operational cost. They wanted an AI agent on top of an existing, mature platform. The real brief was not “add AI.” It was: reduce the analyst workload without creating a new source of risk.

Start with the ground truth

I could not design anything until I understood how the platform worked, where its data lived, what an analyst actually retrieved and which parts of an investigation hurt. Two questions framed the work:

  1. Where in the existing process should an agent go?
  2. How could it connect to legacy systems without transition risk?

These were banks, financial institutions and e-wallets. A hallucination here is not a quirky bug; it is a liability. So the design narrowed around speed to useful evidence and a safe, inspectable path to get there.

Six decisions that kept it trustworthy

Trigger only when work is real

Pre-running AI on every alert would spend tokens on cases that analysts bulk-close and never open. The agent therefore stays asleep until an analyst opens a case. A high-score pre-run can be a later, measured exception.

Make the budget a gate

A fixed five-stage pipeline makes cost knowable before a contract is signed: roughly 40,000 tokens per case, with one visible retry. A free-running loop might make five calls or fifty. Monitoring after the fact is not cost control.

Treat access control as product logic

Every retrieval is row-scoped, personally identifiable information is masked, and reads are logged. Otherwise the friendliest interface can still become a data leak.

Sequence the knowledge, rather than waiting for perfection

Phase one can ship using vendor knowledge and engine output. Phase two builds a client-owned knowledge base from real dispositions. Phase three adds sensitive case history only when trust permits. The client must know that sequence before signing.

Give the client an update path

Typologies, thresholds and procedures change. The client edits its own SOP and skill files through a governed UI. Every revision is versioned and pinned to the case that used it, with a review gate before changes go live.

Cite a source or suppress the statement

Each fact in the summary must trace to a source record and pass schema checks. The model recommends; the analyst accepts, edits or rejects. That feedback can improve later rule mining without disguising a model opinion as ground truth.

None of this was prompt engineering. The hard part was judgement: understanding the business, each person in the room, and the risk.

The result is not an autonomous investigator. It is a bounded assistant with ephemeral working memory, traceable evidence, predictable cost and a human decision at the end.